1 / 60

MEMORY OPTIMIZATION

MEMORY OPTIMIZATION. Christer Ericson Sony Computer Entertainment, Santa Monica (christer_ericson@playstation.sony.com). Talk contents 1/2. Problem statement Why “memory optimization?” Brief architecture overview The memory hierarchy Optimizing for (code and) data cache

Audrey
Download Presentation

MEMORY OPTIMIZATION

An Image/Link below is provided (as is) to download presentation Download Policy: Content on the Website is provided to you AS IS for your information and personal use and may not be sold / licensed / shared on other websites without getting consent from its author. Content is provided to you AS IS for your information and personal use only. Download presentation by click this link. While downloading, if for some reason you are not able to download a presentation, the publisher may have deleted the file from their server. During download, if you can't get a presentation, the file might be deleted by the publisher.

E N D

Presentation Transcript


  1. MEMORY OPTIMIZATION Christer Ericson Sony Computer Entertainment, Santa Monica (christer_ericson@playstation.sony.com)

  2. Talk contents 1/2 • Problem statement • Why “memory optimization?” • Brief architecture overview • The memory hierarchy • Optimizing for (code and) data cache • General suggestions • Data structures • Prefetching and preloading • Structure layout • Tree structures • Linearization caching • …

  3. Talk contents 2/2 • … • Aliasing • Abstraction penalty problem • Alias analysis (type-based) • ‘restrict’ pointers • Tips for reducing aliasing

  4. Problem statement • For the last 20-something years… • CPU speeds have increased ~60%/year • Memory speeds only decreased ~10%/year • Gap covered by use of cache memory • Cache is under-exploited • Diminishing returns for larger caches • Inefficient cache use = lower performance • How increase cache utilization? Cache-awareness!

  5. Need more justification? 1/3 Instruction parallelism: SIMD instructions consume data at 2-8 times the rate of normal instructions!

  6. Need more justification? 2/3 Proebsting’s law: Improvements to compiler technology double program performance every ~18 years! Corollary: Don’t expect the compiler to do it for you!

  7. Need more justification? 3/3 • Consoles don’t follow it (as such) • Fixed hardware • 2nd/3rd generation titles must get improvements from somewhere On Moore’s law:

  8. Brief cache review • Caches • Code cache for instructions, data cache for data • Forms a memory hierarchy • Cache lines • Cache divided into cache lines of ~32/64 bytes each • Correct unit in which to count memory accesses • Direct-mapped • For n KB cache, bytes at k, k+n, k+2n, … map to same cache line • N-way set-associative • Logical cache line corresponds to N physical lines • Helps minimize cache line thrashing

  9. The memory hierarchy Roughly: CPU 1 cycle ~1-5 cycles L1 cache L2 cache ~5-20 cycles Main memory ~40-100 cycles

  10. Some cache specs †16K data scratchpad important part of design ‡configurable as 16K 4-way + 16K scratchpad

  11. Foes: 3 C’s of cache misses • Compulsory misses • Unavoidable misses when data read for first time • Capacity misses • Not enough cache space to hold all active data • Too much data accessed inbetween successive use • Conflict misses • Cache thrashing due to data mapping to same cache lines

  12. Friends: Introducing the 3 R’s • Rearrange (code, data) • Change layout to increase spatial locality • Reduce (size, # cache lines read) • Smaller/smarter formats, compression • Reuse (cache lines) • Increase temporal (and spatial) locality

  13. Measuring cache utilization • Profile • CPU performance/event counters • Give memory access statistics • But not access patterns (e.g. stride) • Commercial products • SN Systems’ Tuner, Metrowerks’ CATS, Intel’s VTune • Roll your own • In gcc ‘-p’ option + define _mcount() • Instrument code with calls to logging class • Do back-of-the-envelope comparison • Study the generated code

  14. Code cache optimization 1/2 • Locality • Reorder functions • Manually within file • Reorder object files during linking (order in makefile) • __attribute__ ((section ("xxx"))) in gcc • Adapt coding style • Monolithic functions • Encapsulation/OOP is less code cache friendly • Moving target • Beware various implicit functions (e.g. fptodp)

  15. Code cache optimization 2/2 • Size • Beware: inlining, unrolling, large macros • KISS • Avoid featuritis • Provide multiple copies (also helps locality) • Loop splitting and loop fusion • Compile for size (‘-Os’ in gcc) • Rewrite in asm (where it counts) • Again, study generated code • Build intuition about code generated

  16. Data cache optimization • Lots and lots of stuff… • “Compressing” data • Blocking and strip mining • Padding data to align to cache lines • Plus other things I won’t go into • What I will talk about… • Prefetching and preloading data into cache • Cache-conscious structure layout • Tree data structures • Linearization caching • Memory allocation • Aliasing and “anti-aliasing”

  17. Prefetching and preloading • Software prefetching • Not too early – data may be evicted before use • Not too late – data not fetched in time for use • Greedy • Preloading (pseudo-prefetching) • Hit-under-miss processing

  18. Software prefetching

  19. Greedy prefetching

  20. Preloading (pseudo-prefetch) (NB: This code reads one element beyond the end of the elem array.)

  21. Structures • Cache-conscious layout • Field reordering (usually grouped conceptually) • Hot/cold splitting • Let use decide format • Array of structures • Structures of arrays • Little compiler support • Easier for non-pointer languages (Java) • C/C++: do it yourself

  22. Field reordering Likely accessed together so store them together!

  23. Hot/cold splitting Cold fields: Hot fields: • Allocate all ‘struct S’ from a memory pool • Increases coherence • Prefer array-style allocation • No need for actual pointer to cold fields

  24. Hot/cold splitting

  25. Beware compiler padding Decreasing size! Assuming 4-byte floats, for most compilers sizeof(X) == 40, sizeof(Y) == 40, and sizeof(Z) == 24.

  26. Cache performance analysis • Usage patterns • Activity – indicates hot or cold field • Correlation – basis for field reordering • Logging tool • Access all class members through accessor functions • Manually instrument functions to call Log() function • Log() function… • takes object type + member field as arguments • hash-maps current args to count field accesses • hash-maps current + previous args to track pairwise accesses

  27. Tree data structures • Rearrange nodes • Increase spatial locality • Cache-aware vs. cache-oblivious layouts • Reduce size • Pointer elimination (using implicit pointers) • “Compression” • Quantize values • Store data relative to parent node

  28. Breadth-first order • Pointer-less: Left(n)=2n, Right(n)=2n+1 • Requires storage for complete tree of height H

  29. Depth-first order • Left(n) = n + 1, Right(n) = stored index • Only stores existing nodes

  30. van Emde Boas layout • “Cache-oblivious” • Recursive construction

  31. A compact static k-d tree union KDNode { // leaf, type 11 int32 leafIndex_type; // non-leaf, type 00 = x, // 01 = y, 10 = z-split float splitVal_type; };

  32. Linearization caching • Nothing better than linear data • Best possible spatial locality • Easily prefetchable • So linearize data at runtime! • Fetch data, store linearized in a custom cache • Use it to linearize… • hierarchy traversals • indexed data • other random-access stuff

  33. Memory allocation policy • Don’t allocate from heap, use pools • No block overhead • Keeps data together • Faster too, and no fragmentation • Free ASAP, reuse immediately • Block is likely in cache so reuse its cachelines • First fit, using free list

  34. The curse of aliasing Aliasing is multiple references to the same storage location What is aliasing? Aliasing is also missed opportunities for optimization What value is returned here? Who knows!

  35. The curse of aliasing • What is causing aliasing? • Pointers • Global variables/class members make it worse • What is the problem with aliasing? • Hinders reordering/elimination of loads/stores • Poisoning data cache • Negatively affects instruction scheduling • Hinders common subexpression elimination (CSE), loop-invariant code motion, constant/copy propagation, etc.

  36. How do we do ‘anti-aliasing’? • What can be done about aliasing? • Better languages • Less aliasing, lower abstraction penalty† • Better compilers • Alias analysis such as type-based alias analysis† • Better programmers (aiding the compiler) • That’s you, after the next 20 slides! • Leap of faith • -fno-aliasing † To be defined

  37. Matrix multiplication 1/3 Consider optimizing a 2x2 matrix multiplication: How do we typically optimize it? Right, unrolling!

  38. Matrix multiplication 2/3 Staightforward unrolling results in this: • But wait! There’s a hidden assumption! a is not b or c! • Compiler doesn’t (cannot) know this! • (1) Must refetch b[0][0] and b[0][1] • (2) Must refetch c[0][0] and c[1][0] • (3) Must refetch b[0][0], b[0][1], c[0][0] and c[1][0]

  39. Matrix multiplication 3/3 A correct approach is instead writing it as: Consume inputs… …before producing outputs

  40. Abstraction penalty problem • Higher levels of abstraction have a negative effect on optimization • Code broken into smaller generic subunits • Data and operation hiding • Cannot make local copy of e.g. internal pointers • Cannot hoist constant expressions out of loops • Especially because of aliasing issues

  41. C++ abstraction penalty • Lots of (temporary) objects around • Iterators • Matrix/vector classes • Objects live in heap/stack • Thus subject to aliasing • Makes tracking of current member value very difficult • But tracking required to keep values in registers! • Implicit aliasing through the this pointer • Class members are virtually as bad as global variables

  42. C++ abstraction penalty Pointer members in classes may alias other members: numVals not a local variable! May be aliased by pBuf! Code likely to refetch numVals each iteration!

  43. C++ abstraction penalty We know that aliasing won’t happen, and can manually solve the aliasing issue by writing code as:

  44. C++ abstraction penalty Since pBuf[i] can only alias numVals in the first iteration, a quality compiler can fix this problem by peeling the loop once, turning it into: Q: Does your compiler do this optimization?!

  45. Type-based alias analysis • Some aliasing the compiler can catch • A powerful tool is type-based alias analysis Use language types to disambiguate memory references!

  46. Type-based alias analysis • ANSI C/C++ states that… • Each area of memory can only be associated with one type during its lifetime • Aliasing may only occur between references of the same compatible type • Enables compiler to rule out aliasing between references of non-compatible type • Turned on with –fstrict-aliasing in gcc

  47. Compatibility of C/C++ types • In short… • Types compatible if differing by signed, unsigned, const or volatile • char and unsigned char compatible with any type • Otherwise not compatible • (See standard for full details.)

  48. What TBAA can do for you It can turn this: Possible aliasing between v[i] and *n into this: No aliasing possible so fetch *n once!

  49. What TBAA can also do • Cause obscure bugs in non-conforming code! • Beware especially so-called “type punning” Illegal C/C++ code! Allowed By gcc Required by standard

More Related